Spotify Automatic Playlist Continuation

End-to-end ML study on the Spotify Million Playlist Dataset (MPD): given a playlist with only some tracks known, predict the rest. ES 335 Course project, IIT Gandhinagar.

Full report / write-up

Results

Hybrid model (ALS matrix factorization + popularity fallback for title-only playlists) vs. a popularity-only baseline, evaluated on our own held-out split of the full 1M-playlist dataset (10,000 eval playlists), using the RecSys Challenge 2018 metrics:

Model	R-precision	NDCG	Clicks
Popularity	0.024	0.075	21.2
Hybrid (ALS + popularity fallback)	0.095	0.219	9.1

See the full report for methodology, EDA, and discussion, and LOG.md for the dated project log.

Project layout

src/                  pipeline scripts (see docs/index.md "Reproducing this")
data/raw/             raw MPD JSON slices (gitignored, not tracked)
data/processed/       built Parquet tables + eval splits (gitignored)
results/metrics/      saved metrics JSON per model
results/figures/      saved plots
docs/                 GitHub Pages report source